ECONOMIC VALUE OF ARTIFICIAL INTELLIGENCE IN HEALTHCARE: A SYSTEMATIC REVIEW OF COST-EFFECTIVENESS ANALYSIS HIGHLIGHTING METHODOLOGICAL GAPS
Author(s)
Ishfaq Rashid Teali, PhD1, Katarzyna Kolasa, PhD2, Muhammed Rashid, PhD3, Carl V. Asche, BA, MBA, MSc, PhD3, Shafia Bashir, PhD4, Nathorn Chaiyakunapruk, PharmD, PhD3, Vicky Tiase, RN, PhD5;
1Tufts Medical Center, Postdoctoral Fellow, Boston, MA, USA, 2Kozminski University, Warsaw, Poland, 3University of Utah, Salt Lake City, UT, USA, 4University of Kashmir, Srinagar, India, 5University of Utah School of Medicine; VA, Salt Lake City, UT, USA
1Tufts Medical Center, Postdoctoral Fellow, Boston, MA, USA, 2Kozminski University, Warsaw, Poland, 3University of Utah, Salt Lake City, UT, USA, 4University of Kashmir, Srinagar, India, 5University of Utah School of Medicine; VA, Salt Lake City, UT, USA
OBJECTIVES: AI/ML are increasingly used in healthcare to enhance diagnosis, prognosis, treatment optimization, and resource allocation. Despite expanding clinical adoption, their economic value remains uncertain, and existing CEAs/CUAs vary substantially in rigor and transparency. This systematic review evaluates CEAs of AI/ML interventions versus conventional care and assesses methodological quality using CHEERS-AI and the Drummond checklist.
METHODS: Following PRISMA 2020 guidelines (PROSPERO CRD420250654972), we searched PubMed, Embase, and the Cochrane Library through February 2025 for peer-reviewed CEAs/CUAs of AI/ML interventions. Data on study characteristics, methods, reporting quality, and economic outcomes were extracted using adapted CHEERS-AI domains.
RESULTS: Thirty-four studies (2018-2025) met inclusion criteria, most of which were model-based economic evaluations (n=27), primarily using Markov or hybrid models; seven were trial-based. AI/ML interventions mainly targeted diagnosis or screening (n=26) and were typically implemented as add-on tools integrated into existing clinical workflows rather than as stand-alone systems (n=31). Among studies reporting incremental cost-effectiveness ratios (ICERs) (30/34), 27 studies found AI/ML interventions to be cost-effective under at least one analytic perspective or scenario, with dominance (lower costs with equal or greater effectiveness) reported in 18 studies. Assessment using the CHEERS-AI extension revealed substantial reporting gaps: no study modeled AI learning over time or reported population differences between training and deployment datasets (0/34), and reporting of AI/ML development, testing, and validation was heterogeneous and often incomplete. Drummond checklist assessments indicated moderate-to-high overall methodological quality, with consistent reporting of core economic evaluation elements but persistent gaps in productivity costing, discounting transparency, and selected uncertainty analyses.
CONCLUSIONS: AI/ML interventions are often cost-effective or dominant, but inconsistent reporting and poor adherence to key CHEERS-AI elements, especially learning over time and deployment population differences, limit confidence and policy relevance. Transparent, lifecycle-aware, and equity-sensitive CEAs are needed to inform reimbursement and implementation decisions.
METHODS: Following PRISMA 2020 guidelines (PROSPERO CRD420250654972), we searched PubMed, Embase, and the Cochrane Library through February 2025 for peer-reviewed CEAs/CUAs of AI/ML interventions. Data on study characteristics, methods, reporting quality, and economic outcomes were extracted using adapted CHEERS-AI domains.
RESULTS: Thirty-four studies (2018-2025) met inclusion criteria, most of which were model-based economic evaluations (n=27), primarily using Markov or hybrid models; seven were trial-based. AI/ML interventions mainly targeted diagnosis or screening (n=26) and were typically implemented as add-on tools integrated into existing clinical workflows rather than as stand-alone systems (n=31). Among studies reporting incremental cost-effectiveness ratios (ICERs) (30/34), 27 studies found AI/ML interventions to be cost-effective under at least one analytic perspective or scenario, with dominance (lower costs with equal or greater effectiveness) reported in 18 studies. Assessment using the CHEERS-AI extension revealed substantial reporting gaps: no study modeled AI learning over time or reported population differences between training and deployment datasets (0/34), and reporting of AI/ML development, testing, and validation was heterogeneous and often incomplete. Drummond checklist assessments indicated moderate-to-high overall methodological quality, with consistent reporting of core economic evaluation elements but persistent gaps in productivity costing, discounting transparency, and selected uncertainty analyses.
CONCLUSIONS: AI/ML interventions are often cost-effective or dominant, but inconsistent reporting and poor adherence to key CHEERS-AI elements, especially learning over time and deployment population differences, limit confidence and policy relevance. Transparent, lifecycle-aware, and equity-sensitive CEAs are needed to inform reimbursement and implementation decisions.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
P2
Topic
Economic Evaluation
Disease
STA: Multiple/Other Specialized Treatments